5D Library › 弃单挽回动力学、沉没数据续费棘轮与高客单顾问筛选机制 · English

第二章 折扣钩子与被高估的挽回

买家会学会等待

许多弃单跟进的第二封或第三封附上一张九折券。短期看,它能挽回更多订单;长期看,它告诉买家一件事:先放进购物车、离开、等一等,就会有折扣。这与耐用品垄断的经典问题同构:如果卖方日后会降价,理性的买家就会等 (Coase 1972)。

推导

# Chapter 2. p03: Discount hooks teach buyers to abandon on purpose.
# A cohort of 1,000 repeat buyers makes 12 purchase decisions a year at
# RM 98. Each decision: a "learner" abandons on purpose and waits for the
# coupon; a non-learner buys at once with probability p, otherwise
# abandons and is recovered by the coupon email with probability r.
# Each coupon received turns a non-learner into a learner with
# probability L. Learners recovered by coupon pay (1 - d) x price.
# Compare revenue per year with coupon emails (d = 10%) and with plain
# reminders (d = 0, no learning, lower recovery r0).
import random

N, ROUNDS, PRICE = 1000, 12, 98.0
P_BUY, R_COUPON, R_PLAIN, D = 0.70, 0.25, 0.17, 0.10

def run(coupon, learn, seed=79):
    rng = random.Random(seed)
    learner = [False] * N
    by_round = []
    for _ in range(ROUNDS):
        rev = 0.0
        for i in range(N):
            if rng.random() < P_BUY:
                if not learner[i]:
                    rev += PRICE
                elif rng.random() < 0.95:
                    rev += PRICE * (1 - D)
                continue
            r = R_COUPON if coupon else R_PLAIN
            if rng.random() < r:
                rev += PRICE * ((1 - D) if coupon else 1.0)
                if coupon and rng.random() < learn:
                    learner[i] = True
        by_round.append(rev)
    return by_round, sum(learner)

if __name__ == "__main__":
    plain, _ = run(False, 0.0)
    print("   plain reminders: RM {:,.0f} per year".format(sum(plain)))
    for learn in (0.0, 0.3, 0.6):
        rows, n = run(True, learn)
        print("   coupon, learning {:.1f}: RM {:,.0f} per year;"
              " {} learners".format(learn, sum(rows), n))
        print("     month 1 RM {:,.0f}, month 12 RM {:,.0f}".format(
            rows[0], rows[-1]))
#    plain reminders: RM 882,196 per year
#    coupon, learning 0.0: RM 903,795 per year; 0 learners
#      month 1 RM 75,440, month 12 RM 75,921
#    coupon, learning 0.3: RM 892,270 per year; 250 learners
#      month 1 RM 75,440, month 12 RM 73,402
#    coupon, learning 0.6: RM 880,971 per year; 427 learners
#      month 1 RM 75,440, month 12 RM 73,235

示意模型中,1,000 名回头客每年做 12 次购买决定,每次 RM 98。只发普通提醒时一年收入 RM 882,196。改发九折券而买家不学习时,一年收入 RM 903,795,确实更多;但如果每收到一张券就有 0.3 的概率学会故意弃单,收入降到 RM 892,270,年底有 250 名「学习者」;学习概率为 0.6 时收入 RM 880,971,已低于普通提醒,年底有 427 名学习者。第一个月三种情形的收入都是 RM 75,440;差别要到年底才显现(第十二个月 RM 73,235,对比不学习时的 RM 75,921)。

注

为什么短期测试看不出来。 一次为期两周的 A/B 测试,比较的是第一个月:九折券看起来稳赢。学习效应要在同一批买家重复购买之后才出现,所以折扣钩子的评估必须按买家追踪数月,或者只对首次购买者使用。

有多少挽回本来就会发生

推导

# Chapter 2. p04: How many "recovered" carts were really recovered?
# Attribution usually credits every order placed after a follow-up was
# sent. With the model of p01, returns that would have happened anyway
# (the natural hazard) are mixed in. Per 1,000 abandoners, RM 98 each.
import importlib.util

def load(name):
    spec = importlib.util.spec_from_file_location(name, name + ".py")
    m = importlib.util.module_from_spec(spec)
    spec.loader.exec_module(m)
    return m

p01 = load("p01_return_hazard")

def after_first(times, n=1000):
    # returns after the first message are what attribution credits
    before = p01.returned((), horizon=times[0])
    total = p01.returned(times)
    return n * (total - before)

if __name__ == "__main__":
    base = p01.returned(())
    for times in ((0.25, 24, 72), (2, 24, 72), (12, 24, 72)):
        att = after_first(times)
        inc = 1000 * (p01.returned(times) - base)
        print("   first message at {:>5g} h: attributed {:4.0f} orders,"
              " incremental {:4.0f}".format(times[0], att, inc))
        print("     incremental share {:.2f}; attributed RM {:,.0f},"
              " incremental RM {:,.0f}".format(inc / att, att * 98,
                                               inc * 98))
#    first message at  0.25 h: attributed  364 orders, incremental  155
#      incremental share 0.43; attributed RM 35,672, incremental RM 15,207
#    first message at     2 h: attributed  326 orders, incremental  149
#      incremental share 0.46; attributed RM 31,960, incremental RM 14,563
#    first message at    12 h: attributed  185 orders, incremental  112
#      incremental share 0.61; attributed RM 18,153, incremental RM 11,025

常见的归因方法把跟进发出之后的每一笔订单都算作「挽回」。用第一章的模型:第一封在 0.25 小时发出时,每 1,000 名弃单者中归因计入 364 笔订单,但跟进真正带来的增量只有 155 笔,占 0.43;按 RM 98 计,归因 RM 35,672,增量 RM 15,207。第一封越晚,归因数字越小,增量所占比例反而越高(12 小时发出时为 0.61),因为最急于回来的人已经自己回来了。

假说 · 尚未验证

作者的假说(尚未验证): 对首次购买者附券、对回头客只发普通提醒的跟进,在 12 个月内的收入不低于全部附券的跟进。什么会证明它错: 按买家随机分组追踪 12 个月,全部附券组的收入显著更高。

小结

折扣钩子在短期内有效,却会教会回头客故意弃单;当学习概率达到 0.6 时,一年收入已低于普通提醒。归因挽回把本来就会回来的买家也算了进去;跟进越早,被高估得越多。

参考文献

Coase, Ronald H. 1972. “Durability and Monopoly.” The Journal of Law and Economics 15 (1): 143–49.

Fan, Y. (2026). Cart-Abandonment Recovery Kinetics, Sunk-Data Retention Ratchets and High-Ticket B2B Qualification Mechanisms: Multi-Pulse Follow-Up Timing, Switching-Cost Accumulation and Screening Equilibria [弃单挽回动力学、沉没数据续费棘轮与高客单顾问筛选机制] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090209 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).